Flip the Explorer to the 202609 release

[rdhyee]

🤖 rbotyee (Claude, operated by @rdhyee) — Raymond's intent: close out the 202609 data release by pointing the live Explorer at it (the minimal wrap-up scope). What I did: finished the 9/9 mechanical rename into a reviewable change — new release manifest, hand-edited CANONICAL.md, docs that say which generation is live, and a fix for preload hints under a data-origin override — then checked it in a browser against the hash-verified 202609 files.

What this changes

• Explorer (explorer.qmd) loads the isamples_202609_* files. It also covers the
  release-manifest filename and the search-index path.
• Release manifest: isamples_202609_release_manifest.json is new, generated
  against the production keys with tools/build_release_manifest.py. The Explorer,
  tests/test_data_origin_contract.py and the data-origin-contract workflow path now
  use it. The 202608 manifest stays in _quarto.yml resources, so already-published
  links and cached older renders keep resolving.
• CANONICAL.md, edited by hand:
  • §1 names the 202609 set, with search-index counts: 256 base + 769 hot shard files,
    1,030 objects.
  • §2's suffix history is left as it was; those rows are all 202608 files.
  • §3 marks 202608 superseded and 202609 live.
  • §4 documents the staging/ copies left on the bucket.
• Preload hints: a small head script now adds the three boot-critical preloads, and
  skips them when ?data_base= or localStorage.ISAMPLES_DATA_BASE overrides the
  origin. Before, previewing a release from another prefix fired three prefetches that
  404'd against the default origin. On a local server the prefetch starts about 0.1 s
  later than with static tags (19 ms → 130 ms, one run each).
• Docs: "what the Explorer serves today" notes now say 202609 (README,
  SERIALIZATIONS, how-to-use, data, tutorials index, why_h3, query-spec,
  DATA_PROVENANCE). Also updated: EXPLORER_QUERIES.md, the guided tour's data URLs,
  and the Explorer's own size/index figures (wide 300 → 237 MB; index 852 → 1,030
  objects). Manifest links now point at isamples.org, where the site deploys the file.
  The old data.isamples.org/isamples_202608_release_manifest.json link returned 404.
• Left alone on purpose: 202608 references that are history or pipeline tooling —
  the comment on the one-off _v2 retrofit, FACET_HIERARCHY_PLAN.md,
  SEARCH_INDEX_V1.md, Makefile, scripts/, and the Worker deploy probes. Those
  probes use 202608 files, which stay served.

202609 vs 202608 — what's the same, what isn't

We compared the live 202608 files directly with the 202609 files.

• Identical:
  • H3 clusters at res4/6/8 (38,462 / 112,019 / 176,669 cells): the same cells, with
    the same per-cell sample count, dominant source and source count. Centres differ
    by at most 0.000001°, which is rounding.
  • sample coordinates: 0 of 6,026,242 located samples moved
  • tree facet summaries and the tree cross-filter, row for row
  • located-sample index metadata: 6,026,242 samples, same per-source counts
  • the mineral ∪ soil tree union (333,253)
  • search results for the three ground-truth queries
• Different:
  • Flat single-value material for 41,372 SESAR samples. Examples: mineral →
    biogenic non-organic material (12,584); soil → biogenic non-organic material
    (11,986); sediment → rock (7,144). Flat facet_summaries material counts move to
    match (e.g. soil 32,149 → 13,806), and about 110 facet_cross_filter rows change.
    In each generation, the summaries agree exactly with that generation's
    sample_facets_v4. The flat label is the first non-root material in each
    sample's list; sql_converter: byte-deterministic output; fix narrow edge row_id collisions pqg#26 made those lists keep source order instead of an
    arbitrary order. So 202609's label is stable where 202608's was arbitrary. Whether the
    first-listed material is the best single label has not been assessed. The default
    tree view is unaffected.
  • Place-name capitalisation for 11,114 located samples (OpenContext 8,457,
    SESAR 2,068, Smithsonian 393, GEOME 196), for example UTAH → Utah and
    3 Coconuts trail → 3 Coconuts Trail. All 11,114 are equal case-insensitively;
    no other place-name content changed.
  • The wide file is 236,995,193 B vs 300,303,095 B, with the same 170 row groups
    and ZSTD. It has 144 fewer rows: OpenContext location rows, from deterministic site
    de-duplication (pqg#26). We did not investigate why the file got smaller.

How it was checked

• tools/verify_release.py check against the hash manifest: 1,046 files, 0 problems.
  This was run on the staged copy, and again on the production keys before this PR.
• Local render (quarto render explorer.qmd, Quarto 1.8.24; CI pins 1.6.42), served
  with dev_server.py, loaded in Chrome against the 202609 files:
  • 38,462 clusters, 6,026,242 samples
  • SESAR 4,389,231 / OpenContext 1,104,985 / GEOME 291,210 / Smithsonian 240,816
  • __manifestCheck ok (only extra: offline-only df.parquet)
  • zero console errors
  • search totals: basalt 785, pottery Cyprus 1,305, axial seamount summit caldera 284
• tests/test_data_origin_contract.py passes against the production origin with the
  new manifest.
• Codex review: see the final comment on this PR.

Rollback

Revert this PR's merge commit (git revert -m 1 <merge sha>). The 202608 files on R2
were not modified, so the revert redeploys a fully working 202608 Explorer.

— isamples-0910 (Claude, for rdhyee)

🤖 Generated with Claude Code

https://claude.ai/code/session_012aY9Nvapmp5zMUPATZygb7

[rdhyee]

🤖 rbotyee (Claude, operated by @rdhyee) — Raymond's intent: have this flip reviewed by a second AI before it ships, as with the rest of the 202609 chain. What I did: ran three Codex review rounds over the branch diff and am posting the verdicts.

Round 1 — CHANGES REQUESTED, four P2 findings, all fixed in the branch:

1. CANONICAL.md §3 understated the data change — the 202609 row now states that content is not identical and names both differences (flat material for 41,372 SESAR samples; 144 fewer location rows), plus what is unchanged.
2. DATA_PROVENANCE.md carried reproducibility/version-skew statements that read as claims about 202609 — now scoped to the historical 202601/202604 builds, and flagged as not re-audited for the 202609 re-run.
3. Deleting isamples_202608_release_manifest.json would have broken its published URL — the file is restored unchanged and both manifests stay in _quarto.yml resources.
4. The recorded start of 202608's production life was wrong — corrected to the Production: OC sync (#272) + #277/#283 fixes → isamples_202608 #284 cutover (merged 2026-06-14), matching commit d26fabb.

Round 2 — LGTM. All four verified resolved; no new merge-blocking findings; 111 Node unit tests pass; runtime filenames, manifest entries and builder inventory agree. Codex noted its approval was conditional on the production promotion, the manifest regeneration and passing contract checks — all three have since been done and are described in the PR body.

Round 3 — LGTM, no findings. A narrow round over the later CANONICAL.md §3 edit (place-name capitalisation and the precise H3 wording). Codex reproduced the counts from the parquet files independently, and confirmed that the case-only place-name changes produce identical search-index token sequences, so search results are unaffected.

Claude + Codex: LGTM. Not merged — Raymond's call.

— isamples-0910 (Claude, for rdhyee)